What the system can detect or produce
The task is defined by the output the application needs. Several tasks can be combined when one image must answer different questions.
Classification
Assign one or more labels to an image or cropped region.
label + confidence
Object detection
Locate and classify each relevant object in a frame.
boxes + labels
Segmentation
Identify the pixels belonging to an object, surface, or condition.
masks + area
Tracking
Follow the same object through video and record its movement or state.
track + events
OCR and extraction
Read text and map fields from documents, labels, screens, or scenes.
text + fields
Visual inspection
Find defects, missing parts, unexpected patterns, or process deviations.
finding + location
Examples with a defined decision
Manufacturing inspection
Check whether a required part is present, locate a surface defect, and send uncertain cases to an operator.
Warehouse and logistics
Read labels, count items, detect loading events, or identify damaged packages from fixed cameras or handheld images.
Document processing
Classify document pages, locate tables or fields, read their contents, and preserve page coordinates for review.
Site and equipment monitoring
Detect defined conditions in periodic images or video and create an event only when duration and confidence rules are met.
The test set must look like the real environment
A model can score well on clean images and fail after installation. Evaluation data must cover the conditions that change what the camera sees.
Light
Day, night, shadows, reflections, glare, and seasonal changes.
View
Camera height, angle, distance, lens, crop, and resolution.
Scene
Occlusion, clutter, similar objects, motion blur, and background changes.
Frequency
Common cases, rare faults, new variants, and periods with no target event.
Hardware
Camera compression, frame rate, connectivity, compute, and installation limits.
People
Privacy zones, retention, access, review responsibilities, and lawful use.
Images, annotation, and model choice
Existing cameras and sample media are inspected before the model is selected. The review shows whether image quality supports the task and which conditions are missing from the available data.
Annotation guidelines define exactly what counts as an object, defect, boundary, or event. A sample is labelled by more than one reviewer to find ambiguous rules before the full dataset is prepared.
A pretrained model or managed vision API is used when it meets the requirements. Custom training is useful when the target, environment, or error tolerance is specific enough that general models do not perform reliably.
Accuracy is tied to the decision
The right threshold depends on what happens after a prediction. Missing a safety event and sending an extra item for review do not have the same cost.
- Detection quality
- Precision, recall, localization, class confusion, and performance by site or condition.
- Decision threshold
- The confidence level and duration rules that create, suppress, or route an event.
- System performance
- End-to-end latency, frame rate, uptime, bandwidth, and hardware use.
- Field review
- A monitored trial records missed cases, false alerts, camera changes, and operator feedback.
Where inference can run
Cloud
Suitable when connectivity is stable and centralized scaling or managed hardware is useful.
Private infrastructure
Keeps media and inference inside a company-controlled network or cloud account.
Edge device
Processes frames near the camera when latency, bandwidth, offline operation, or data minimization matters.
Hybrid
Runs immediate detection at the edge and sends selected events or metadata for central review and storage.
What you receive
- •Task and camera feasibility assessment
- •Annotation guide and versioned datasets
- •Model or vision API integration
- •Evaluation set, thresholds, and test report
- •Inference service for the agreed environment
- •Application integration and review workflow
- •Monitoring, retraining triggers, and operating documentation
Computer vision development questions
Can existing cameras be used?
Often, but a sample review is needed. Resolution, lens, placement, lighting, compression, and frame rate must make the target visible at the required level of detail.
Do we need to label our own images?
A pretrained system may cover common objects or OCR. A company-specific defect, product, scene, or event usually needs representative labelled examples. The initial pilot determines how much additional data is useful.
Can it work in real time?
Yes, if the model, hardware, resolution, frame rate, and application latency are designed together. Some tasks need every frame; others can sample frames or process only a triggered event.
How are false alerts handled?
Thresholds are set against the cost of missed and extra detections. Low-confidence cases can be grouped, delayed until an event persists, or sent for human review instead of creating an automatic action.
Can images stay on site?
Yes. Inference can run on an edge device or private server, and the system can retain only selected events or structured results. The exact design depends on hardware and audit requirements.